Post on 05-Oct-2020
Abstract—This paper proposes an application of human skin
region detection using Hybrid Particle Swarm Optimization
(HPSO) algorithm. It consists of two steps. In the first step, the
input RGB color image is converted into CIEL*a*b color space.
Then this is clustered by the Hillclimbing segmentation with
K-Means clustering algorithm, which will be useful to find the
number of clusters and the local optimal solutions. In the second
step, these local solutions are further improved by PSO
algorithm using YCbCr explicit skin color conditions in order
to find the global solution. This solution helps to detect the
robust skin region. Finally the performance measure named
Peak Signal-to-Noise Ratio (PSNR) is performed on the
ground-truth skin dataset. The experimental results has shown
the efficiency of the proposed method.
Index Terms—3D histogram, color space, hillclimbing
segmentation with k-means, particle swarm optimization,
PSNR.
I. INTRODUCTION
Applications of image segmentation can be found in a
wide variety of areas such as remote sensing, vehicle and
robot navigation, medical imaging, surveillance, target
identification and tracking, scene analysis, product
inspection/quality control etc. Image segmentation remains a
long-standing problem in computer vision and has been
found difficult and challenging for two main reasons: firstly,
the fundamental complexity of modelling a vast amount of
visual data that appear in the image is a considerable
challenge; the second challenge is the intrinsic ambiguity in
image perception, especially when it concerns so-called
unsupervised segmentation. Hence, the proposed application
focused on the cluster based segmentation which
automatically determines the “optimum” number of clusters
to the application of human skin region detection.
Skin region detection plays an important role in many
applications of image processing. The applications are face
detection, face recognition, gesture recognition, human body
parts like hand, eye and lip detection and others. The
important cue to detect the skin region is by skin color. So,
color spaces play a major role to detect the skin region.
Different color spaces like RGB, Normalized RGB, HSV,
YCbCr [1],[2],[3],[4],[5] and CIEL*a*b [6] are mainly used.
Since many recent proposals are based on the underlying idea
of representing the skin color in an optimal color space by
means of the so-called skin cluster. Color information is an
Manuscript received June 20, 2012; revised August 5, 2012.
R. Vijayanandh is with the Development Centre, Bharathiar University,
Coimbatore, Tamil Nadu, India (e-mail: rvanandh@gmail.com).
G. Balakrishnan is with the Department of Computer Science and
Engineering, Indra Ganesan College of Engineering, Tiruchirappalli, Tamil
Nadu, India (e-mail: balakrishnan.g@gmail.com).
efficient tool for identifying skin areas if the skin color model
can be properly adapted for different lighting environments.
This fact leads to avoid the use of RGB, because the red,
green and blue components are highly correlated and
dependent on lighting conditions. Hence, the proposed
method is based on CIEL*a*b and YCbCr color spaces.
Some of the skin color modeling techniques are skin
detection rule, parametric skin distribution modeling,
non-parametric skin distribution modelling [2] and edge
based [4].
Optimization technique is very important in image
segmentation applications. Since, this helps to minimize
certain criteria (e.g., intra cluster centre). The optimization
techniques are Simulated Annealing, Genetic Algorithms,
Ant Colony Optimization algorithms and Particle Swarm
Optimization. In that, PSO has been applied to many
different applications of segmentation [1],[7],[8],[9], pattern
classification and image analysis.
This proposed paper deals with 3D Histogram, dynamic
cluster formation, local and global optimum solution. The
local optimum solutions are determined by Hillclimbing
segmentation with K-Means clustering algorithm of
CIEL*a*b color image. Then these local solutions are further
refined by the PSO algorithm, in order to find the global
solution by YCbCr color space explicit skin color conditions.
The paper is structured as follows. In Section 2, literature
review discussed. The Section 3, explain the formation of
clusters and finding local solutions. The Section 4 describes
the proposed HPSO algorithm. The Section 5 and 6,
illustrates the performance measure and demonstrates the
experimental results respectively. The conclusion and future
work is discussed in the final section.
II. LITERATURE REVIEW
J. A. Nasiri, H. S. Yazdi and M. Naghibzadeh [1] proposed
the chrominance based mouth detection approach using PSO
rule mining. The high values of Cr and low values of Cb
considered as the mouth region. PSO had been tuned to
separate pixels properly.
Nils Janssen and Neil Robertson [2] used a non-parametric
classification scheme based on a histogram similarity
measure by various color spaces and the resolution of
histograms. Also compared the results with three methods
namely Gaussian, Bayesian and Thresholding.
A. Cheddad, J. Condell, K. Curran and P. Mc Kevitt [3] set
hypothesis that “luminance inclusion does increase
separability of skin and non-skin clusters”. This approach
used a new color space which contains error signals derived
from differentiating the grayscale map and the
Performance Measure of Human Skin Region Detection
Based on Hybrid Particle Swarm Optimization
R. Vijayanandh, Member, IACSIT and G. Balakrishnan
857
International Journal of Computer Theory and Engineering, Vol. 4, No. 5, October 2012
non-encoded-red grayscale version and performed the
reduction of space from 3D to 1D.
A.Y. Dawod, J. Abdullah and Md. J. Alam [4] also
proposed the luminance based clustering. Edge detection
could be used to construct a free-form skin color model.
Gradient method by Sobel operator detected edges at finest
scales and has smoothing action along the edge direction.
Qing-Fang Zheng and Wen Gao [5] proposed a fast
adaptive skin detection approach that worked on DCT
domain of JPEG image and classifies each image block
according to its color and texture properties. The color and
texture features adopted by YCbCr color space, since it is
more valid than other color spaces for skin detection. The
experiment carried out in two ways: first, the accuracy and
efficiency of the method were compared with existing
adaptive methods and non-adaptive method. Second, the
advantage of using both color and texture features.
D. D. Gomez, C. Butakoff, B. K. Ersboll and W. Stoecker
[6] proposed an unsupervised algorithm named Independent
Histogram Pursuit for segmenting dermatological lesions.
This method was used to find a suitable image-dependent
linear transformation of an arbitrary multispectral color space
to aid segmentation of dermatological images. The spectral
bands was enhanced the contrast between healthy skin and
lesion. Then the remaining N-1 combinations were estimated.
Chih-Cheng Hung and Li Wan [7] proposed the image
classification by hybridization of Particle Swarm
Optimization with the K-Means algorithm. This method was
performed in two steps: in the first step, different PSO
heuristics was optimized by the K-Means algorithm. In the
second step, determined the reliability parametric values for
different variants of PSO and K-Means algorithms.
Mahamed G.H. Omran, A. Salman and A. P. Engelbrecht
[8] proposed a new dynamic clustering approach based on
particle swarm optimization. Binary particle swarm
optimization was used to automatically determine the
optimum number of clusters. Then the chosen clusters were
refined by K-Means clustering algorithm. The experiments
were conducted by both synthetic images and natural images.
Yuhua Zheng and Yan Meng [9] proposed a robust
tracking algorithm used an adaptive tracking window
associated with five parameters. These parameters were
optimized by a PSO algorithm. The fitness function for
particles was calculated by appearance histogram.
The proposed paper used the concepts of YCbCr color
space, CIEL*a*b color space, 3D Histogram, Hillclimbing
K-Means clustering algorithm and PSO for finding a global
solution. The above existing works clearly explained the
importance of these concepts. The drawbacks identified by
the existing works could be optimally reduced through
combining these concepts. and also this is a good solution for
the drawbacks mentioned in our previous works [10] and
[11].
III. CLUSTER FORMATION AND FINDING LOCAL SOLUTIONS
K-Means clustering algorithm is one of the simplest and
efficient unsupervised learning algorithms for clustering
problems such as image segmentation. Such efficient
algorithm could be used here for identifying the number of
clusters and the local solutions.
The input color image is first converted into CIEL*a*b
color space. Since CIEL*a*b color space is human
perception based color space. 3D histogram and histogram
size (number of bins) (here 10) are used to find the histogram
peaks. This is used to obtain the number of initial seeds for
K-Means clustering algorithm. The results of K-Means
clustering algorithm and the number of cluster are passed to
PSO algorithm for further refinement, in order to obtain the
global best solution. 3D histograms assume dependence of
one channel on the other and which is more realistic. This
could be the same implementation as in third color space
discussed in [12] and [13]. Fig. 1 shows the sample labelled
segmentation results of Hillclimbing segmentation with
K-Means clustering algorithm. Fig. 2 shows the distribution
of pixels in CIEL*a*b color space of the Fig. 1 labelled
images.
(a) (b)
(c) (d)
Fig. 1. Segmentation results of Hillclimbing segmentation with
K-Means clustering algorithm (a), (c) Original image (b), (d) Segmented
labelled image.
Fig. 2. Distribution of pixels in CIEL*a*b color space of Fig. 1 labelled
images.
Table I shows the performance measures of automatic
cluster formation by Hillclimbing segmentation with
K-Means clustering algorithm.
TABLE I: CLUSTER FORMATION BY HILLCLIMBING SEGMENTATION WITH
K-MEANS CLUSTER ALGORITHM
Input Image Image
Size
Number
of
Clusters
Duration
(seconds)
Fig 1 (a) 170 x 235 4 0.689
Fig 1 (c) 450 x 600 4 4.796
858
International Journal of Computer Theory and Engineering, Vol. 4, No. 5, October 2012
IV. PROPOSED HYBRID PARTICLE SWARM OPTIMIZATION
ALGORITHM
Particle Swarm Optimization was originally introduced by
Kennedy and Eberhart [14], inspired by social behaviour of
bird flocking or fish schooling.
In PSO, a swarm of individuals (called particles) fly
through the search space. Each particle represents a candidate
solution to the optimization problem. The position of a
particle is influenced by the best position visited by itself and
the position of the best particle in its neighborhood. When the
neighborhood of a particle is the entire swarm, the best
position in the neighborhood is referred to as the global best
particle and the resulting algorithm is referred to as a gbest
PSO. When smaller neighborhoods are used, the algorithm is
generally referred to as a lbest PSO. The performance of each
particle is measured using a fitness function that varies
depending on the optimization problem [8]. It is a population
based stochastic optimization technique in which individuals
called particles change their positions and velocity with time.
Also PSO is easy to implement for its simple particle
updating operation. It is a new method which can solve most
of global optimization problems.
Each particle in the swarm is represented by the following
characteristics:
idx : The current position of the particle.
idv : The current velocity of the particle.
idp : The personal best position of the particle.
The new velocity and position of each particle can be
calculated using the current velocity and the distance from
ip to gp as shown in the following formulas:
)(())(() 21
1 k
idgd
k
idid
k
id
k
id xprandcxprandcvwv
1 1 (2)k k k
id id idx x v
where .,...,2,1,,...,2,1 mdni ,0,, 21 ccw n is the
number of particles in a group, m is the number of members
in a particle, w is the inertia weight factor, 1c and 2c are
acceleration constants, rand are random numbers with the
range 1,0 . The particles follow up the scent ip and gp to
search in the goal searching space until the optimum result
does not change again or reach the scheduled numbers.
PSO algorithm is used to optimize the objective functions
that are mainly used to minimize within cluster distance
(between pixel and the cluster mean) and to maximize among
cluster distance (between cluster pairs). In this proposed
paper, the number of clusters is not known priori. This is
derived from the Hillclimbing segmentation with K-Means
clustering algorithm and the details of the cluster formation is
discussed in the previous section.
The detailed procedure for combining the K-Means
clustering algorithm and PSO algorithm is outlined as
follows:
1) The input RGB color image is converted into CIEL*a*b
color space.
2) The CIEL*a*b color space components undergone the
calculation of histogram in 3D, in order to find the
histogram peaks.
3) This peak value is used as the number of clusters of
K-Means clustering algorithm.
4) The return values, the number of cluster and cluster
centres of K-Means clustering algorithm are passed to
the proposed PSO algorithm.
5) The PSO algorithm uses the return values of K-Means
and take the argument as RGB color image as in step 1.
Here this original image can be converted into YCbCr
color space.
6) The PSO algorithm used to find the global optimum
solution from the initial solutions of K-Means clustering
algorithm with the following explicit skin color
conditions and prerequisites:
Population is derived from Hillclimbing segmentation with K-Means clustering algorithm.
Maximum and minimum chrominance values are assigned to Cb=125 & Cr=172 and Cb=75 & Cr=132 respectively.
The solutions of K-Means clustering algorithm are assigned to the best values of the proposed method.
The number of iteration is assigned to 100. Individual weight and social weight of the particles is
2 each. Inertial factor (w) is 1. PSO algorithm stores the chrominance components
previous best position of each particle and the best global position of particles by using equations (1) and (2).
7) The cluster of the skin region mapped to the original
RGB components of the input image and other clusters
are eradicated. Here all non-skin region color
components are assigned to zero.
Finally the segmented image is converted into gray scale
image and performed the performance measure with PSNR
by using MCG skin ground-truth dataset.
V. PERFORMANCE MEASURE
The performance measurement for image segmentation,
the well known Peak Signal-to-Noise Ratio (in dB) which is
classified under the difference distortion metrics is applied on
the skin region detected by the proposed method and the
ground-truth of skin images. It is defined as:
10
25520log ( ) ( ) (3)PSNR dB
RMSE
where RMSE is the root mean-squared error, defined as:
MxN
jiIjiI
RMSE
M
i
N
j
1 1
2)),(ˆ),((
Here I and I are original and segmented images of size
is MxN respectively.
PSNR is often expressed on a logarithmic scale in decibels
(dB). PSNR values falling below 30 dB indicate a fairly low
quality. However, a high quality image should strive for
(1)
859
International Journal of Computer Theory and Engineering, Vol. 4, No. 5, October 2012
40dB and above [15]. The PSNR value is calculated between
the gray converted image of the proposed segmented image
and the ground-truth of skin image [16].
VI. EXPERIMENTAL RESULTS
This proposed method was implemented by Matlab7.0 on
a 2.4GHz Pentium IV machine running on 256MB RAM.
Experiments were conducted by the images of MCG skin
ground-truth dataset, different country people and various
illuminations, totally 385 images. These images have
different numbers of clusters with varying complexities. The
proposed PSO algorithm parameters are empirically set as
follows: Nc is derived from the number of clusters of
K-Means clustering algorithm and w=1, c1=c2=2.
Fig. 1 (b) and Fig. 1 (d) shows the labelled cluster image
formed by Hillclimbing segmentation with K-Means
clustering algorithm, which is same as discussed in [12] and
[13]. This clustering algorithm could be useful to find the
number of cluster and to find the local optimal solution of the
input images.
Then these solutions are further refined by the proposed
PSO algorithm by YCbCr color space explicit skin color
conditions. The skin region cluster identified by the proposed
method, which is explained in the step 6 of the previous
section. The experiments on Hillclimbing segmentation with
K-Means clustering algorithm and PSO algorithm are
intended to show the effectiveness of the results of the
classification. Finally the performance measure of PSNR is
calculated between the gray converted segmented result and
MCG skin ground-truth dataset. The experiment could be
carried out by the MCG ground-truth skin dataset, skin
different illumination and lighting condition images, people
from various country and small skin patch images. Fig. 3 and
Fig. 4 show the skin region pixels distribution of the Fig. 1
Original images and the experimental result of the proposed
method respectively.
Fig. 3. Skin region pixels distribution of the proposed method of Fig. 1
Original images.
Hence the results are more robust and also overcome the
drawbacks mentioned in our previous works [10] and [11].
Fig. 4 (d), (f) and (l) are the best sample results, to prove the
proposed method is more robust than the existing work.
(a) (b)
(c) (d)
(e) (f)
(g) (h)
(i) (j)
(k) (l)
Fig. 4. Experimental results of the proposed PSO algorithm (a), (c), (e), (g),
(i), (k) Original image (b), (d), (f), (h), (j), (l) The proposed PSO algorithm
based skin region segmented image.
The PSNR values calculated by the proposed method for
Fig. 4 (b) and (d) is illustrated in Table II and the
860
International Journal of Computer Theory and Engineering, Vol. 4, No. 5, October 2012
corresponding skin ground-truth images is shown in Fig. 5 (b)
and Fig. 5 (d). TABLE II: PSNR VALUES.
Input Image PSNR
Value
Fig 5 (a) 41.14
Fig 5 (c) 41.14
(a) (b)
(c) (d)
Fig. 5. Skin ground-truth image (a), (c) Original image (b), (d) MCG skin
ground-truth image.
VII. CONCLUSION AND FUTURE WORK
In this paper, a new skin region detection algorithm based
on PSO was proposed. Dynamic formation of clusters by
Hillclimbing segmentation with K-Means clustering
algorithm and the chrominance based global solution of the
skin cluster found by PSO algorithm makes the process more
robust to detect the skin region. The performance measure
and experimental results show that our method is satisfactory
for skin region detection. These satisfactory results were
based on the segmentation quality, duration of execution,
dynamic cluster formation, less frame buffer usage and
optimal global solution. Also the segmented results will be
useful to many computer vision applications like Face
detection, Face recognition, Gesture recognition and etc.
Future work will extend the experiments on low-resolution
images with texture and spatial information instead of the
chrominance values.
REFERENCES
[1] J. A. Nasiri, H. S. Yazdi, and M. Naghibzadeh, “A mouth detection
approach based on PSO rule mining on color images,” 5th Iranian
Conference on Machine Vision and Image Processing, November 4-6,
2008.
[2] N. Janssen and N. Robertson, “On the detection of low-resolution skin
regions in surveillance images,” The Eighth International Workshop on
Visual Surveillance, 2008.
[3] A. Cheddad, J. Condell, K. Curran, and P. Mc Kevitt, “A new colour
space for skin tone detection,” ICIP 2009, pp. 497-500.
[4] A. Y. Dawod, J. Abdullah, and Md. J. Alam, “A new method for hand
segmentation using free-form skin color model,” 3rd International
Conference on Advanced Computer Theory and Engineering, 2010,
pp.V2-562 – V2-566.
[5] Q.-F. Zheng and W. Gao, “Fast adaptive skin detection in JPEG
images,” PCM 2005, Part II, LNCS 3768, pp. 595-605.
[6] D. D. Gomez, C. Butakoff, B. K. Ersboll, and W. Stoecker,
“Independent histogram pursuit for segmentation of skin lesions,”
IEEE Transactions on Biomedical Engineering, vol. 55, no. 1, January
2008, pp. 157-161.
[7] C.-C. Hung and W. Li, “Hybridization of particle swarm optimization
with the k-means algorithm for image classification,” ISBN
978-1-4244-2760-4/09, 2009.
[8] M. G. H. Omran, A. Salman, and A. P. Engelbrecht, “Dynamic
clustering using particle swarm optimization with application in image
segmentation,” Pattern Analysis Application, 2006, vol. 8, pp.
332-344.
[9] Y. H. Zheng and M. Yan, “The PSO-based adaptive window for people
tracking,” Proceedings of the 2007 IEEE Symposium on Conputational
Intelligence in Security and Defense Applications, 2007, pp. 23-29.
[10] R. Vijayanandh and G. Balakrishnan, “A hybrid approach to human
skin region detection,” ICTACT Journal on Image and Video
Processing, February 2011, vol. 1, Issue 3, pp. 143 - 149.
[11] R. Vijayanandh and G. Balakrishnan, “Skin region detection using
hillclimbing segmentation with fuzzy c-means algorithm,”
International Journal of Advanced Engineering and Applications, vol.
II, Jan’2011, pp. 200 – 204.
[12] R. Vijayanandh and G. Balakrishnan, “Human skin region
identification using fusion technique,” Springer-Verlag Berlin
Heidelberg, ICT 2010, CCIS 101, pp. 622-625.
[13] R. Vijayanandh and G. Balakrishnan, “Human face detection using
color spaces and region property measures,” Proceedings of the 11th
International Conference on Control, Automation, Robotics and Vision
(ICARCV 2010), 2010, pp. 1605-1610.
[14] J. Kennedy and R. Eberhart, “Particle swarm optimization,”
Proceedings of IEEE International Conference on Neural Networks,
Perth, 1995.
[15] A. Cheddad, J. Condell, K. Curran, and P. M. Kevitt, “Skin tone based
streganography in video files exploiting the YCbCr colour space,”
ICME 2008, pp. 905 – 908.
[16] H. Lei, T. Xiz, Y. D. Zhang, and S. X. Lin, “Human skin detection in
images by MSER analysis,” 2011 18th International Conference on
Image Processing, pp. 1281-1284, 2011.
R. Vijayanandh received his B.Sc., Mathematics
(Silver Medal) degree from St. Joseph’s College,
Tiruchirappalli, Tamil Nadu, India in 1997. Also he
obtained M.C.A and M.Phil (Computer Science)
degrees from Bharathidasan University,
Tiruchirappalli, Tamil Nadu, India in 2000 and 2005
respectively. He is currently pursuing Ph.D (Computer
Science) at Bharathiar University, Coimbatore, Tamil
Nadu, India.
He has 12 years of academic experience. Currently he is working as
Assistant Professor and Head in the Department of Computer Applications,
Imayam College of Information Technology, Tamil Nadu, India. He has
published and presented many research papers in the International Journals
and International Conferences held in India and abroad. He also received a
travel grant from AICTE, New Delhi, India to present his paper in Nanyang
Technological University, Singapore.
R. Vijayanandh is a member of IAENG, ISTE, IACSIT and AIRCC. His
primary research interest includes Image Segmentation, Computer Vision
and Vision-based Computer Interaction.
Dr. G. Balakrishnan received his Bachelors and
Masters degree in Computer Science and Engineering
from Bharathidasan University and Bharathiar
University respectively. He received the PhD degree
from University Malaysia Sabah, Malaysia in 2006.
He is the Director of Indra Ganesan College of
Engineering, Tiruchirappalli, Tamil Nadu, India. He
has 10 years of academic experience. He has published
more than 30 technical papers in various International
Journals and Conferences. Dr. G. Balakrishnan is the active member of ISTE,
CSI and IAENG. His research interest includes Image Processing, Neural
Networks and Fuzzy Logic.
861
International Journal of Computer Theory and Engineering, Vol. 4, No. 5, October 2012